A Functional Space Model. COS 326 David Walker Princeton University

Transcription

1 A Functional Space Model COS 326 David Walker Princeton University

2 Data type representations: Last Time type tree = Leaf Node of int * tree * tree Leaf: Node(i, left, right): 0 Node 3 left right

3 This Time Understanding the space complexity of functional programs At least two interesting components: the amount of live space at any instant in time the rate of allocation a function call may not change the amount of live space by much but may allocate at a substantial rate because functional programs act by generating new data structures and discarding old ones, they often allocate at a great rate» OCaml garbage collector is optimized with this in mind» interesting fact: at the assembly level, the number of writes made by a function program is typically roughly the same as the number of writes by an imperative program

4 This Time Understanding the space complexity of functional programs At least two interesting components: the amount of live space at any instant in time the rate of allocation a function call may not change the amount of live space by much but may allocate at a substantial rate because functional programs act by generating new data structures and discarding old ones, they often allocate at a great rate» OCaml garbage collector is optimized with this in mind» interesting fact: at the assembly level, the number of writes made by a function program is typically roughly the same as the number of writes by an imperative program What takes up space? conventional first-order data: tuples, lists, strings, datatypes function representations (closures) the call stack

5 CONVENTIONAL DATA

6 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i)

7 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

8 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

9 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

10 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

11 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

12 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

13 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

14 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Consider: insert t 21 t

15 Allocating space Whenever you use a constructor, space is allocated: let rec insert (t:tree) (i:int) = match t with Leaf -> Node (i, Leaf, Leaf) Node (j, left, right) -> if i <= j then Node (j, insert left i, right) else Node (j, left, insert right i) Total space allocated is proportional to the height of the tree. ~ log n, if tree with n nodes is balanced t

16 Compare let check_option (o:int option) : int option = match o with Some _ -> o None -> failwith found none let check_option (o:int option) : int option = match o with Some j -> Some j None -> failwith found none

17 Compare let check_option (o:int option) : int option = match o with Some _ -> o None -> failwith found none allocates nothing when arg is Some i let check_option (o:int option) : int option = match o with Some j -> Some j None -> failwith found none allocates an option when arg is Some i

18 Compare let cadd (c1:int*int) (c2:int*int) : int*int = let (x1,y1) = c1 in let (x2,y2) = c2 in (x1+x2, y1+y2) let double (c1:int*int) : int*int = let c2 = c1 in cadd c1 c2 let double (c1:int*int) : int*int = cadd c1 c1 let double (c1:int*int) : int*int = let (x1,y1) = c1 in cadd (x1,y1) (x1,y1)

19 Compare let cadd (c1:int*int) (c2:int*int) : int*int = let (x1,y1) = c1 in let (x2,y2) = c2 in (x1+x2, y1+y2) let double (c1:int*int) : int*int = let c2 = c1 in cadd c1 c2 c1 1 2 c2 let double (c1:int*int) : int*int = cadd c1 c1 let double (c1:int*int) : int*int = let (x1,y1) = c1 in cadd (x1,y1) (x1,y1)

20 Compare let cadd (c1:int*int) (c2:int*int) : int*int = let (x1,y1) = c1 in let (x2,y2) = c2 in (x1+x2, y1+y2) let double (c1:int*int) : int*int = let c2 = c1 in cadd c1 c2 let double (c1:int*int) : int*int = cadd c1 c1 c1 1 2 let double (c1:int*int) : int*int = let (x1,y1) = c1 in cadd (x1,y1) (x1,y1)

21 Compare let cadd (c1:int*int) (c2:int*int) : int*int = let (x1,y1) = c1 in let (x2,y2) = c2 in (x1+x2, y1+y2) let double (c1:int*int) : int*int = let c2 = c1 in cadd c1 c2 let double (c1:int*int) : int*int = cadd c1 c1 let double (c1:int*int) : int*int = let (x1,y1) = c1 in cadd (x1,y1) (x1,y1) c1 1 2 arg1 1 2 arg2 1 2

22 Compare let cadd (c1:int*int) (c2:int*int) : int*int = let (x1,y1) = c1 in let (x2,y2) = c2 in (x1+x2, y1+y2) let double (c1:int*int) : int*int = let c2 = c1 in cadd c1 c2 cadd allocates double does not let double (c1:int*int) : int*int = cadd c1 c1 cadd allocates double does not let double (c1:int*int) : int*int = let (x1,y1) = c1 in cadd (x1,y1) (x1,y1) cadd allocates double allocates 2 pairs

23 Compare let cadd (c1:int*int) (c2:int*int) : int*int = let (x1,y1) = c1 in let (x2,y2) = c2 in (x1+x2, y1+y2) let double (c1:int*int) : int*int = let (x1,y1) = c1 in cadd c1 c1 cadd allocates double does not extracts components; does not allocate

24 FUNCTION CLOSURES

25 Consider the following program: Closures let choose (arg:bool * int * int) : int -> int = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2)

26 Consider the following program: Closures let choose (arg:bool * int * int) : int -> int = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) It s execution behavior according to the substitution model: choose (true, 1, 2)

27 Consider the following program: Closures let choose (arg:bool * int * int) : int -> int = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) It s execution behavior according to the substitution model: choose (true, 1, 2) let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y)

28 Consider the following program: Closures let choose (arg:bool * int * int) : int -> int = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) It s execution behavior according to the substitution model: choose (true, 1, 2) let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y) if true then (fun n -> n + 1) else (fun n -> n + 2)

29 Consider the following program: Closures let choose (arg:bool * int * int) : int -> int = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) It s execution behavior according to the substitution model: choose (true, 1, 2) let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y) if true then (fun n -> n + 1) else (fun n -> n + 2) (fun n -> n + 1)

30 let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) Substitution and Compiled Code

31 Substitution and Compiled Code let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) compile choose: mov rb r_arg[0] mov rx r_arg[4] mov ry r_arg[8] compare rb 0... jmp ret main:... jmp choose

32 Substitution and Compiled Code let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) execute with substitution let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y) compile choose: mov rb r_arg[0] mov rx r_arg[4] mov ry r_arg[8] compare rb 0... jmp ret main:... jmp choose

33 Substitution and Compiled Code let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) execute with substitution let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y) compile choose: mov rb r_arg[0] mov rx r_arg[4] mov ry r_arg[8] compare rb 0... jmp ret main:... jmp choose execute with substitution == generate new code block with parameters replaced by arguments

34 Substitution and Compiled Code let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) execute with substitution let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y) compile choose: mov rb r_arg[0] choose_subst: mov rx r_arg[4] mov rb 0xF8[0] mov ry r_arg[8]... jmp ret main:... jmp choose choose: mov rb r_arg[0] mov rx r_arg[4] mov ry r_arg[8] compare rb 0... jmp ret main:... jmp choose execute with substitution == generate new code block with parameters replaced by arguments mov rx 0xF8[4] mov ry 0xF8[8] compare rb 0... jmp ret 0xF8: 0 1 2

35 Substitution and Compiled Code let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y) choose (true, 1, 2) execute with substitution let (b, x, y) = (true, 1, 2) in if b then (fun n -> n + x) else (fun n -> n + y) execute with substitution if true then (fun n -> n + 1) else (fun n -> n + 2) compile choose: mov rb r_arg[0] mov rx r_arg[4] mov ry r_arg[8] compare rb 0... jmp ret main:... jmp choose execute with substitution == generate new code block with parameters replaced by arguments choose: mov rb r_arg[0] choose_subst: mov rx r_arg[4] mov rb 0xF8[0] mov ry r_arg[8] mov rx 0xFF44]... choose_subst2: mov ry 0xFF84[8] jmp ret compare 1 0 compare rb main: jmp ret jmp ret... jmp choose 0xF8: 0 1 2

36 What we aren t going to do The substitution model of evaluation is just a model. It says that we generate new code at each step of a computation. We don t do that in reality. Too expensive! The substitution model is a faithful model for reasoning about program correctness but it doesn t help us understand what is going on at the machine-code level that s a good thing! abstraction!! you should almost never think about machine code when writing a program. We invented high-level programming languages so you don t have to. Still, we need to have a more faithful space model in order to understand how to write efficient algorithms.

37 Some functions are easy to implement let add (x:int*int) : int = let (y,z) = x in y + z # argument in r1 # return address in r0 add: ld r2, r1[0] ld r3, r1[4] add r4, r2, r3 jmp r0 # y in r2 # z in r3 # sum in r4 If no functions in ML were nested then compiling ML would be just like compiling C. (Take COS 320 to find out how to do that...)

38 How do we implement functions? Let s remove the nesting and compile them like we compile C. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y)??? let choose arg = let (b, x, y) = arg in if b then f1 else f2 let f1 n = n + x let f2 n = n + y

39 How do we implement functions? Let s remove the nesting and compile them like we compile C. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x) else (fun n -> n + y)??? let choose arg = let (b, x, y) = arg in if b then f1 else f2 let f1 n = n + x let f2 n = n + y Darn! Doesn t work naively. Nested functions contain free variables. Simple unnesting leaves them undefined.

40 How do we implement functions? We can t define a function like the following using code alone: let f2 n = n + y A closure is a pair of some code and an environment: code let f2 (n,env) = n + env.y {y = 1} environment closure

41 Closure Conversion Closure conversion (also called lambda lifting) converts open, nested functions in to closed, top-level functions. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x + y) else (fun n -> n + y)

42 Closure Conversion Closure conversion (also called lambda lifting) converts open, nested functions in to closed, top-level functions. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x + y) else (fun n -> n + y) let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) add environment parameter create closures let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye use environment variables instead of free variables

43 Closure Conversion Closure conversion (also called lambda lifting) converts open, nested functions in to closed, top-level functions. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x + y) else (fun n -> n + y) let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) add environment parameter create closures (choose (true,1,2)) 3 let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye use environment variables instead of free variables let c_closure = (choose, ()) in (* create closure *) let (c_code, c_cenv) = c_closure in (* extract code, env *) let (f_code, f_env) = c_code ((true,1,2), c_env) in (* call choose code, extract f code, env *) f_code (3, f_env) (* call f code *)

44 Closure Conversion Closure conversion (also called lambda lifting) converts open, nested functions in to closed, top-level functions. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x + y) else (fun n -> n + y) let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) add environment parameter create closures (choose (true,1,2)) 3 let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye use environment variables instead of free variables let c_closure = (choose, ()) in (* create closure *) let (c_code, c_cenv) = c_closure in (* extract code, env *) let (f_code, f_env) = c_code ((true,1,2), c_env) in (* call choose code, extract f code, env *) f_code (3, f_env) (* call f code *)

45 Closure Conversion Closure conversion (also called lambda lifting) converts open, nested functions in to closed, top-level functions. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x + y) else (fun n -> n + y) let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) add environment parameter create closures (choose (true,1,2)) 3 let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye use environment variables instead of free variables let c_closure = (choose, ()) in (* create closure *) let (c_code, c_cenv) = c_closure in (* extract code, env *) let (f_code, f_env) = c_code ((true,1,2), c_env) in (* call choose code, extract f code, env *) f_code (3, f_env) (* call f code *)

46 Closure Conversion Closure conversion (also called lambda lifting) converts open, nested functions in to closed, top-level functions. let choose arg = let (b, x, y) = arg in if b then (fun n -> n + x + y) else (fun n -> n + y) let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) add environment parameter create closures (choose (true,1,2)) 3 let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye use environment variables instead of free variables let c_closure = (choose, ()) in (* create closure *) let (c_code, c_cenv) = c_closure in (* extract code, env *) let (f_code, f_env) = c_code ((true,1,2), c_env) in (* call choose code, extract f code, env *) f_code (3, f_env) (* call f code *)

47 One Extra Note: Typing Even though the original, non-closure-converted code was welltyped, the closure-converted code isn t because the environments are different let choose (arg,env) = let (b, x, y) = arg in if b then (f1, F1 {xe=x; ye=y}) else (f2, F2 {ye=y}) let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye type f1_env = {x1:int; y1:int} type f2_env = {y2:int} type f1_clos = (int * f1_env -> int) * f1_env type f2_clos = (int * f2_env -> int) * f2_env

48 One Extra Note: Typing Even though the original, non-closure-converted code was welltyped, the closure-converted code isn t because the environments are different let choose (arg,env) = let (b, x, y) = arg in if b then (f1, F1 {x1=x; y2=y}) else (f2, F2 {y2=y}) type f1_env = {x1:int; y1:int} type f2_env = {y2:int} let f1 (n,env) = match env with F1 e -> n + e.x1 + e.y2 F2 _ -> failwith "bad env!" let f2 (n,env) = match env with F1 _ -> failwith "bad env!" F2 e -> n + e.y2 type f1_clos = (int * f1_env -> int) * f1_env type f2_clos = (int * f2_env -> int) * f2_env fix I: type env = F1 of f1_env F2 of f2_env type f1_clos = (int * env -> int) * env type f2_clos = (int * env -> int) * env

49 One Extra Note: Typing Even though the original, non-closure-converted code was welltyped, the closure-converted code isn t because the environments are different let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye type f1_env = {xe:int; ye:int} type f2_env = {xe:int} type f1_clos = (int * f1_env -> int) * f1_env type f2_clos = (int * f2_env -> int) * f2_env fix II: type f1_env = {xe:int; ye:int} type f2_env = {xe:int} type f1_clos = exists env.(int * env -> int) * env type f2_clos = exists env.(int * env -> int) * env

50 One Extra Note: Typing Even though the original, non-closure-converted code was welltyped, the closure-converted code isn t because the environments are different let choose (arg,env) = let (b, x, y) = arg in if b then (f1, {xe=x; ye=y}) else (f2, {ye=y}) let f1 (n,env) = n + env.xe + env.ye let f2 (n,env) = n + env.ye From System F to Typed Assembly Language, -- Morrisett, Walker et al. type f1_env = {xe:int; ye:int} type f2_env = {xe:int} type f1_clos = (int * f1_env -> int) * f1_env type f2_clos = (int * f2_env -> int) * f2_env fix II: type f1_env = {xe:int; ye:int} type f2_env = {xe:int} type f1_clos = exists env.(int * env -> int) * env type f2_clos = exists env.(int * env -> int) * env

51 map has a universal polymorphic type: Aside: Existential Types map : ('a -> 'b) -> 'a list -> 'b list "for all types 'a and for all types 'b, " when we closure-convert a function that has type int -> int, we get a function with existential polymorphic type: exists 'a. ((int * 'a) -> int) * 'a "there exists some type 'a such that, " In OCaml, we can approximate existential types using datatypes (a data type allows you to say "there exists a type 'a drawn from one of the following finite number of options." In Haskell, you've got the real thing.

52 Closure Conversion: Summary (before) (after) All function definitions equipped with extra env parameter: let f arg =... let f_code (arg, env) =... All free variables obtained from environment: x env.cx All functions values paired with environment: f (f_code, {ve1=v1;...; ven=vn}) All function calls extract code and environment and call code: f e let (f_code, f_env) = f in f_code (e, f_env)

53 The Space Cost of Closures The space cost of a closure = the cost of the pair of code and environment pointers + the cost of the data referred to by function free variables

54 TAIL CALLS AND CONTINUATIONS

55 Some Innocuous Code (* sum of 0..n *) let rec sum_to (n:int) : int = if n > 0 then n + sum_to (n-1) else 0 let big_int = sum big_int Let s try it. (Go to tail.ml)

56 Some Other Code Four functions: Green works on big inputs; Red doesn t. let sum_to2 (n: int) : int = let rec aux (n:int) (a:int) : int = if n > 0 then aux (n-1) (a+n) else a in aux n 0 let rec sum_to (n:int) : int = if n > 0 then n + sum_to (n-1) else 0 let rec sum2 (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum2 tail let sum (l:int list) : int = let rec aux (l:int list) (a:int) : int = match l with [] -> a hd::tail -> aux tail (a+hd) in aux l 0

57 Some Other Code Four functions: Green works on big inputs; Red doesn t. let sum_to2 (n: int) : int = let rec aux (n:int) (a:int) : int = if n > 0 then aux (n-1) (a+n) else a in aux n 0 let rec sum_to (n:int) : int = if n > 0 then n + sum_to (n-1) else 0 code that works: no computation after recursive function call let rec sum2 (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum2 tail let sum (l:int list) : int = let rec aux (l:int list) (a:int) : int = match l with [] -> a hd::tail -> aux tail (a+hd) in aux l 0

58 Tail Recursion A tail-recursive function is a function that does no work after it calls itself recursively. (* sum of 0..n *) let sum_to2 (n: int) : int = let rec aux (n:int)(a:int) : int = if n > 0 then aux (n-1) (a+n) else a in aux n 0

59 Tail Recursion A tail-recursive function is a function that does no work after it calls itself recursively. Tail-recursive: sum_to aux aux aux aux 0 ( ) (* sum of 0..n *) let sum_to2 (n: int) : int = let rec aux (n:int)(a:int) : int = if n > 0 then aux (n-1) (a+n) else a in aux n 0 constant size expression (addition overflow occurred at some point)

60 Tail Recursion A tail-recursive function is a function that does no work after it calls itself recursively. Not tail-recursive: sum_to sum_to sum_to sum_to add it all back up... (* sum of 0..n *) let rec sum_to (n:int) : int = if n > 0 then n + sum_to (n-1) else 0 let big_int = sum big_int expression grows at every recursive call

61 Memory is partitioned: Stack and Heap heap space (big!) stack space (small!)

62 Data Needed on Return Saved on Stack sum_to the stack every non-tail call puts the data from the calling context on the stack not much space left! will run out soon!

63 Question Can any non-tail-recursive function be transformed in to a tailrecursive one? let rec sum_to (n: int) : int = if n > 0 then n + sum_to (n-1) else 0 human ingenuity not only is sum2 tail-recursive but it reimplements an algorithm that took linear space (on the stack) using an algorithm that executes in constant space! let sum_to2 (n: int) : int = let rec aux (n:int)(a:int) : int = if n > 0 then aux (n-1) (a+n) else a in aux n 0

64 Question Can any non-tail-recursive function be transformed in to a tailrecursive one? Yes, if we can capture the differential between a tail-recursive function and a non-tail-recursive one. let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail Idea: Focus on what happens after the recursive call.

65 Question Can any non-tail-recursive function be transformed in to a tailrecursive one? Yes, if we can capture the differential between a tail-recursive function and a non-tail-recursive one. let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail what happens next Idea: Focus on what happens after the recursive call. Extracting that piece: hd + How do we capture it?

66 Question How do we capture that computation? hd + fun s -> hd + s

67 Question How do we capture that computation? hd + fun s -> hd + s let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s ->???)

68 Question How do we capture that computation? hd + fun s -> hd + s let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s))

69 Question How do we capture that computation? hd + fun s -> hd + s let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int =??

70 Question How do we capture that computation? hd + fun s -> hd + s let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s)

71 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2]

72 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s)

73 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s))

74 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s)) sum_cont [] (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s))

75 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s)) sum_cont [] (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) 0

76 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s)) sum_cont [] (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) 0 (fun s -> (fun s -> s) (1 + s)) (2 + 0))

77 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s)) sum_cont [] (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) 0 (fun s -> (fun s -> s) (1 + s)) (2 + 0)) (fun s -> s) (1 + (2 + 0))

78 Execution type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s)) sum_cont [] (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s)) 0 (fun s -> (fun s -> s) (1 + s)) (2 + 0)) (fun s -> s) (1 + (2 + 0)) 1 + (2 + 0) 3

79 Question type cont = int -> int let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum (l:int list) : int = sum_cont l (fun s -> s) sum [1;2] sum_cont [1;2] (fun s -> s) sum_cont [2] (fun s -> (fun s -> s) (1 + s)) sum_cont [] (fun s -> (fun s -> (fun s -> s) (1 + s)) (2 + s))... 3 Where did the stack space go?

80 CPS CPS: short for Continuation-Passing Style Every function takes a continuation as an argument that expresses "what to do next" CPS functions only call other functions as the last thing they do

81 CORRECTNESS OF A CPS TRANSFORM

82 type cont = int -> int Are the two functions the same? let rec sum_cont (l:int list) (k:cont): int = match l with [] -> k 0 hd::tail -> sum_cont tail (fun s -> k (hd + s)) let sum2 (l:int list) : int = sum_cont l (fun s -> s) let rec sum (l:int list) : int = match l with [] -> 0 hd::tail -> hd + sum tail Here, it is really pretty tricky to be sure you've done it right if you don't prove it. Let's try to prove this theorem and see what happens: for all l:int list, sum_cont l (fun x => x) == sum l

83 Attempting a Proof for all l:int list, sum_cont l (fun s -> s) == sum l Proof: By induction on the structure of the list l. case l = []... case: hd::tail IH: sum_cont tail (fun s -> s) == sum tail

84 Attempting a Proof for all l:int list, sum_cont l (fun s -> s) == sum l Proof: By induction on the structure of the list l. case l = []... case: hd::tail IH: sum_cont tail (fun s -> s) == sum tail == sum_cont (hd::tail) (fun s -> s)

85 Attempting a Proof for all l:int list, sum_cont l (fun s -> s) == sum l Proof: By induction on the structure of the list l. case l = []... case: hd::tail IH: sum_cont tail (fun s -> s) == sum tail sum_cont (hd::tail) (fun s -> s) == sum_cont tail (fn s' -> (fn s -> s) (hd + s')) (eval)

86 Attempting a Proof for all l:int list, sum_cont l (fun s -> s) == sum l Proof: By induction on the structure of the list l. case l = []... case: hd::tail IH: sum_cont tail (fun s -> s) == sum tail sum_cont (hd::tail) (fun s -> s) == sum_cont tail (fn s' -> (fn s -> s) (hd + s')) (eval) == sum_cont tail (fn s' -> hd + s') (eval -- hd + s' valuable)

87 Need to Generalize the Theorem and IH for all l:int list, sum_cont l (fun s -> s) == sum l Proof: By induction on the structure of the list l. case l = []... case: hd::tail IH: sum_cont tail (fun s -> s) == sum tail sum_cont (hd::tail) (fun s -> s) == sum_cont tail (fn s' -> (fn s -> s) (hd + s')) (eval) == sum_cont tail (fn s' -> hd + s') (eval -- hd + s' valuable) == darn! we'd like to use the IH, but we can't! we might like: not the identity continuation (fun s -> s) like the IH requires sum_cont tail (fn s' -> hd + s') == sum tail... but that's not even true

88 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l)

89 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] must prove: for all k:int->int, sum_cont [] k == k (sum [])

90 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] must prove: for all k:int->int, sum_cont [] k == k (sum []) pick an arbitrary k:

91 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] must prove: for all k:int->int, sum_cont [] k == k (sum []) pick an arbitrary k: sum_cont [] k

92 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] must prove: for all k:int->int, sum_cont [] k == k (sum []) pick an arbitrary k: sum_cont [] k == match [] with [] -> k 0 hd::tail ->... (eval) == k 0 (eval)

93 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] must prove: for all k:int->int, sum_cont [] k == k (sum []) pick an arbitrary k: sum_cont [] k == match [] with [] -> k 0 hd::tail ->... (eval) == k 0 (eval) == k (sum [])

94 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] must prove: for all k:int->int, sum_cont [] k == k (sum []) pick an arbitrary k: sum_cont [] k == match [] with [] -> k 0 hd::tail ->... (eval) == k 0 (eval) == k (0) (eval, reverse) == k (match [] with [] -> 0 hd::tail ->...) (eval, reverse) == k (sum []) case done!

95 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] ===> done! case l = hd::tail IH: for all k':int->int, sum_cont tail k' == k' (sum tail) Must prove: for all k:int->int, sum_cont (hd::tail) k == k (sum (hd::tail))

96 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] ===> done! case l = hd::tail IH: for all k':int->int, sum_cont tail k' == k' (sum tail) Must prove: for all k:int->int, sum_cont (hd::tail) k == k (sum (hd::tail)) Pick an arbitrary k, sum_cont (hd::tail) k

97 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] ===> done! case l = hd::tail IH: for all k':int->int, sum_cont tail k' == k' (sum tail) Must prove: for all k:int->int, sum_cont (hd::tail) k == k (sum (hd::tail)) Pick an arbitrary k, sum_cont (hd::tail) k == sum_cont tail (fun s -> k (hd + x)) (eval)

98 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] ===> done! case l = hd::tail IH: for all k':int->int, sum_cont tail k' == k' (sum tail) Must prove: for all k:int->int, sum_cont (hd::tail) k == k (sum (hd::tail)) Pick an arbitrary k, sum_cont (hd::tail) k == sum_cont tail (fun s -> k (hd + x)) (eval) == (fun s -> k (hd + s)) (sum tail) (IH with IH quantifier k' replaced with (fun x -> k (hd+x))

99 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] ===> done! case l = hd::tail IH: for all k':int->int, sum_cont tail k' == k' (sum tail) Must prove: for all k:int->int, sum_cont (hd::tail) k == k (sum (hd::tail)) Pick an arbitrary k, sum_cont (hd::tail) k == sum_cont tail (fun s -> k (hd + x)) (eval) == (fun s -> k (hd + s)) (sum tail) (IH with IH quantifier k' replaced with (fun x -> k (hd+x)) == k (hd + (sum tail)) (eval, since sum total and and sum tail valuable)

100 Need to Generalize the Theorem and IH for all l:int list, for all k:int->int, sum_cont l k == k (sum l) Proof: By induction on the structure of the list l. case l = [] ===> done! case l = hd::tail IH: for all k':int->int, sum_cont tail k' == k' (sum tail) Must prove: for all k:int->int, sum_cont (hd::tail) k == k (sum (hd::tail)) Pick an arbitrary k, sum_cont (hd::tail) k == sum_cont tail (fun s -> k (hd + x)) (eval) == (fun s -> k (hd + s)) (sum tail) (IH with IH quantifier k' replaced with (fun x -> k (hd+x)) == k (hd + (sum tail)) (eval, since sum total and and sum tail valuable) == k (sum (hd:tail)) (eval sum, reverse) case done! QED!

101 Finishing Up Ok, now what we have is a proof of this theorem: for all l:int list, for all k:int->int, sum_cont l k == k (sum l) We can use that general theorem to get what we really want: for all l:int list, sum2 l == sum_cont l (fun s -> s) (by eval sum2) == (fun s -> s) (sum l) (by theorem, instantiating k with (fun s -> s) == sum l So, we've show that the function sum2, which is tail-recursive, is functionally equivalent to the non-tail-recursive function sum.

102 SUMMARY

103 Summary of the CPS Proof We tried to prove the specific theorem we wanted: for all l:int list, sum_cont l (fun s -> s) == sum l But it didn't work because in the middle of the proof, the IH didn't apply -- inside our function we had the wrong kind of continuation -- not (fun s -> s) like our IH required. So we had to prove a more general theorem about all continuations. for all l:int list, for all k:int->int, sum_cont l k == k (sum l) This is a common occurrence -- generalizing the induction hypothesis -- and it requires human ingenuity. It's why proving theorems is hard. It's also why writing programs is hard -- you have to make the proofs and programs work more generally, around every iteration of a loop.

104 Overall Summary We developed techniques for reasoning about the space costs of functional programs the cost of manipulating data types like tuples and trees the cost of allocating and using function closures the cost of tail-recursive and non-tail-recursive functions We also talked about some important program transformations: closure conversion makes nested functions with free variables in to pairs of closed code and environment the continuation-passing style (CPS) transformation turns non-tailrecursive functions in to tail-recursive ones that use no stack space the stack gets moved in to the function closure since stack space is often small compared with heap space, it is often necessary to use continuations and tail recursion but full CPS-converted programs are unreadable: use judgement

105 Challenge: CPS Convert the incr function type tree = Leaf Node of int * tree * tree let rec incr (t:tree) (i:int) : tree = match t with Leaf -> Leaf Node (j,left,right) -> Node (i+j, incr left i, incr right i) (see solution after the next slide)

106 END

107 CPS Convert the incr function type tree = Leaf Node of int * tree * tree let rec incr (t:tree) (i:int) : tree = match t with Leaf -> Leaf Node (j,left,right) -> Node (i+j, incr left i, incr right i) type cont = tree -> tree let rec incr_cps (t:tree) (i:int) (k:cont) : tree = match t with Leaf -> k Leaf Node (j,left,right) ->...

108 type tree = Leaf Node of int * tree * tree let rec incr (t:tree) (i:int) : tree = match t with Leaf -> Leaf Node (j,left,right) -> Node (i+j, incr left i, incr right i) first continuation: Node (i+j,, incr right i) second continuation: Node (i+j, left_done, )

109 type tree = Leaf Node of int * tree * tree let rec incr (t:tree) (i:int) : tree = match t with Leaf -> Leaf Node (j,left,right) -> Node (i+j, incr i left, incr i right) first continuation: fun left_done -> Node (i+j, left_done, incr right i) second continuation: fun right_done -> k (Node (i+j, left_done, right_done))

110 type tree = Leaf Node of int * tree * tree let rec incr (t:tree) (i:int) : tree = match t with Leaf -> Leaf Node (j,left,right) -> Node (i+j, incr left i, incr right i) second continuation inside first continuation: fun left_done -> let k2 = (fun right_done -> k (Node (i+j, left_done, right_done)) ) in incr right i k2

111 type tree = Leaf Node of int * tree * tree let rec incr (t:tree) (i:int) : tree = match t with Leaf -> Leaf Node (j,left,right) -> Node (i+j, incr left i, incr right i) type cont = tree -> tree let rec incr_cps (t:tree) (i:int) (k:cont) : tree = match t with Leaf -> k Leaf Node (j,left,right) -> let k1 = (fun left_done -> let k2 = (fun right_done -> k (Node (i+j, left_done, right_done))) in incr_cps right i k2 ) in incr_cps left i k1 let incr_tail (t:tree) (i:int) : tree = incr_cps t i (fun t -> t)